Uncertainty-aware probabilistic graph neural networks for road-level traffic crash prediction.

Area of Science:

• Urban planning and transportation safety
• Artificial intelligence and machine learning
• Statistical modeling and risk assessment

Background:

• Traffic crashes pose significant risks to urban safety and mobility.
• Existing prediction models struggle with fine spatiotemporal scales and inherent crash uncertainties.
• Current methods often fail to provide a nuanced understanding of crash risk hierarchies.

Purpose of the Study:

• To develop an uncertainty-aware probabilistic deep learning model for road-level, daily traffic crash prediction.
• To address limitations of traditional methods in capturing sporadic crash events and non-crash data.
• To provide a comprehensive prediction of crash risk, including high, low, and no-risk scenarios.

Main Methods:

• Introduction of the Spatiotemporal Zero-Inflated Tweedie Graph Neural Networks (STZITD-GNN).
• Integration of the Tweedie statistical family for non-Gaussian crash data with graph neural networks.
• Utilization of a zero-inflated component to distinguish non-crash and low-risk situations.

Main Results:

• The STZITD-GNN model demonstrated superior performance over baseline models in real-world data from London, UK.
• Achieved up to a 34.60% reduction in regression error for point estimation.
• Improved interval-based uncertainty metrics by over 47%.

Conclusions:

• The STZITD-GNN is the first uncertainty-aware probabilistic graph deep learning model for multi-step, road-level traffic crash prediction.
• The model offers a holistic view of road safety by accurately predicting and differentiating various crash risk levels.
• This approach enhances urban mobility safety by providing more precise and reliable traffic crash risk insights.